CN103640532A - Pedestrian anti-collision early warning method based on recognition of braking and accelerating intention of driver - Google Patents

Abstract

The invention discloses a pedestrian anti-collision early warning method based on identification of the driver's braking and acceleration intentions, which comprises the following steps: 1. collecting experimental data for an off-line training of a hidden Markov HMM model; ii. collecting experimental data for hidden Secondary off-line training of the Markov HMM model; Ⅲ. Identify the data signal of the driver's collision avoidance intention and pedestrian's intention; Different alert handling. In the present invention, pedestrians and vehicles are studied as a system, and by analyzing the manipulation behavior and strategy that the driver may adopt when encountering pedestrians in front of the vehicle, and judging whether there is danger according to the driving behavior and intention, the driver is given a corresponding early warning , to give early warning to wrong driver's driving operation such as stepping on the accelerator pedal by mistake, effectively protecting the safety of pedestrians and improving the active safety performance of the car.

Description

Pedestrian anti-collision warning method based on recognition of driver's braking and acceleration intentions

technical field

The invention relates to the field of automobile active safety, in particular to a pedestrian anti-collision early warning method based on recognition of driver's braking and acceleration intentions.

Background technique

In road traffic accidents, pedestrians are often the biggest victim group, and collisions between cars and pedestrians are one of the main types of accidents. According to the U.S. Highway Traffic Safety Administration statistics, in 2011, 69,000 pedestrians were injured due to traffic accidents in the United States, accounting for 3% of the total number of injuries; 4,432 pedestrians were killed, accounting for 14% of the total number of deaths. In road traffic accidents in the European Union, the death data of pedestrians is 9 times that of occupants in vehicles, and the death data of cyclists is 8 times that of occupants in vehicles. In 2010, 16,281 pedestrians were killed and 44,629 were injured due to traffic accidents in my country, accounting for 25% and 18% of the total respectively. At present, pedestrian protection has attracted worldwide attention. Strict collision standards and pedestrian protection regulations have been formulated in the passive safety of automobiles. In the active safety of automobiles, sensor technology is used to sense the pedestrians in front and judge their dangerous status, and timely warn the driver that the vehicle may be in contact with the pedestrian. Pedestrians in front are in danger of collision and realize active safety warning.

The current research on driver behavior mainly focuses on the detection and supervision of a certain dangerous driving behavior, without using environmental information to comprehensively consider the driver's driving behavior and intention, ignoring the driver's intention and its changing trend in active safety control. It is easy to make a wrong estimate of the current road danger situation. The present invention intends to identify and predict the driver's driving behavior and intention by using the hidden Markov model based on the pedestrian detection results, combined with the vehicle's operating status information and surrounding environment information, and analyze the possible actions taken by the driver when encountering pedestrians in front of the vehicle. According to the driving behavior and intention whether there is danger, the driver and the pedestrian in front will be given corresponding warnings, effectively protecting the safety of pedestrians and improving the active safety performance of the car. .

Contents of the invention

In view of the defects in the prior art, the purpose of the present invention is to provide a pedestrian anti-collision warning method based on driver braking and acceleration intention recognition, by analyzing the driver's possible maneuvering behavior when encountering a pedestrian in front of the vehicle and strategies, and judge whether there is danger based on driving behavior and intentions, give corresponding warnings to drivers, and give warnings to wrong driver’s driving operations such as wrongly stepping on the accelerator pedal, effectively protecting the safety of pedestrians and improving the active safety of cars performance.

In order to achieve the above object, technical scheme of the present invention:

A pedestrian anti-collision warning method based on driver braking and acceleration intention identification includes the following steps:

ⅰ. Collect experimental data for an offline training of the hidden Markov HMM model:

According to the driving behavior that the driver may take to avoid collision with pedestrians, collect experimental data, the experimental data includes brake pedal force data, brake pedal displacement data, accelerator pedal travel data and vehicle speed data; After the data is segmented and processed, the brake pedal force data, brake pedal displacement data, and accelerator pedal travel data are input into the braking and acceleration hidden Markov HMM model, and the vehicle speed data is input into the speed classification module; In the hidden Markov HMM model of acceleration, a total of 9 models are constructed, including normal release of the accelerator, quick release of the accelerator, accelerator hold, step on the accelerator, normal step on the brake, fast step on the brake, brake hold, release of the brake and no action on the pedal. A multidimensional Gaussian hidden Markov HMM model about braking and acceleration; apply Baum-Welch algorithm, carry out offline training to described 9 multidimensional Gaussian hidden Markov HMM models, optimize each model parameter iteratively; The vehicle speed signal is numbered according to the grade and input to the speed classification module;

ii. Collect experimental data for secondary offline training of the hidden Markov HMM model: collect experimental data for the driving behavior that the driver may take to avoid collisions with pedestrians. The experimental data includes brake pedal force data, brake pedal Displacement data, accelerator pedal stroke data, and vehicle speed data; input brake pedal force data, brake pedal displacement data, and accelerator pedal stroke data into the braking and acceleration hidden Markov HMM model again, and input vehicle speed data into the speed In the classification module; apply the Forward-Backward algorithm to calculate the likelihood of the newly collected driving behavior experimental data relative to the 9 driving behavior multidimensional Gaussian hidden Markov HMM models described in step i, select the maximum likelihood The model is used as the driver's driving behavior identification result; and the two-dimensional driving behavior identification result string of the Cain Markov HMM model-braking and acceleration identification result string and vehicle speed identification result string is used as the driver's intention identification hidden Markov The observation sequence of the HMM model is used for offline training and optimization of the driver collision avoidance pedestrian intention hidden Markov model, and two driver collision avoidance pedestrian intention hidden Markov HMM models are obtained: Accelerated hidden Markov HMM model and system Parking Hidden Markov HMM model;

ⅲ. Identify the data signal of the driver’s collision avoidance pedestrian intention: input the real-time collected brake pedal force sensor signal, brake pedal displacement sensor signal, and accelerator pedal stroke sensor signal into the braking and acceleration hidden Markov HMM model, Input the vehicle speed sensor signal into the speed classification module, conduct off-line training and optimization for 9 driving behavior multi-dimensional Gaussian hidden Markov HMM models, identify the driver's operation, and obtain the two-dimensional identification of the driving behavior hidden Markov HMM model Result string—braking and acceleration identification result string and vehicle speed identification result string, after composing the observation sequence string, send it into two hidden Markov HMM models for driver intention identification, and apply the Forward-Backward algorithm to calculate two multi-dimensional discrete hidden HMM models respectively. The Markov HMM model generates the possibility of the observation sequence, and the model with the largest likelihood is selected as the driving intention data signal;

ⅳ. After identifying the data signal of the driver’s intention to avoid collision with pedestrians, analyze the data combined with the infrared human body sensing data signal, and make different early warning processing: install a human body sensor based on infrared technology at the front of the car for real-time collection of human body sensing Data signal: a central processing unit is set in the vehicle-mounted system for real-time data analysis based on the human body sensing data signal and driving intention data signal and controls the sound signal warning device for early warning.

The data analysis process includes: if the central processing unit analyzes the human body sensing data signal and the result is that there is no pedestrian in front, then no matter what the result of the driving intention data signal is, the central processing unit will not send a trigger signal to control the sound signal warning device for early warning ; If the central processing unit analyzes the result of the human body sensing data signal that there are pedestrians ahead and the driving intention data signal is a brake parking intention signal; the central processing unit does not send a trigger signal to control the sound signal warning device for early warning; if the central processing unit analyzes The result of the human body sensing data signal is that there is a pedestrian in front and the driving intention data signal is an acceleration intention signal, that is, when the driver accidentally steps on the accelerator pedal to accelerate and pass, the central processing unit sends a trigger signal to control the sound signal warning device to give an early warning, so that the driver can Respond in time to protect the purpose of pedestrians.

Compared with prior art, the beneficial effect of the present invention:

The present invention studies pedestrians and vehicles as a system, and provides a driver's braking and acceleration intention identification method applied to pedestrian anti-collision warnings. By analyzing the driver's possible manipulation behavior and According to the driving behavior and intention to judge whether there is danger, give corresponding early warning to the driver, and give early warning to the wrong driver's driving operation such as the operation of stepping on the accelerator pedal by mistake, effectively protecting the safety of pedestrians and improving the active safety performance of the car .

Description of drawings

Fig. 1 is a block diagram of the overall design scheme of the pedestrian anti-collision warning method based on driver's braking and acceleration intention recognition in the present invention.

Fig. 2 is the HMM model structure of the pedestrian anti-collision warning method based on driver's braking and acceleration intention recognition in the present invention.

Fig. 3 is the training process of the HMM model of the pedestrian anti-collision warning method based on driver's braking and acceleration intention recognition in the present invention.

FIG. 4 is a flow chart of the pedestrian anti-collision warning method based on driver's braking and acceleration intention recognition according to the present invention.

In the figure: 1. Human sensor, 2. Brake pedal travel sensor, 3. Brake pedal force sensor, 4. Accelerator pedal travel sensor, 5. Vehicle speed sensor, 6. Hidden Markov model module, 7. Central processing Unit, 8, sound signal warning device.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings.

As shown in Figure 1, the design idea of the present invention is described by constructing an example of a pedestrian anti-collision warning method based on driver braking and acceleration intention recognition. husband model module 6, central processing unit 7, sound signal warning device 8 and the vehicle information collection device installed in the automobile; the vehicle information collection device includes: brake pedal travel sensor 2, brake pedal force sensor 3, accelerator pedal Stroke sensor 4 and vehicle speed sensor 5; Wherein said brake pedal stroke sensor 2, brake pedal force sensor 3, accelerator pedal stroke sensor 4, vehicle speed sensor 5 are respectively connected with Hidden Markov Model module 6 signals; The device 1, the hidden Markov model module 6 and the sound signal warning device 8 are respectively connected with the central processing unit 7.

Wherein, the sound signal warning device 8 can use a buzzer to sound.

Wherein, the collection of data is realized by the vehicle information collection device, and the vehicle information collection device includes a brake pedal stroke sensor 2, a brake pedal force sensor 3, an accelerator pedal stroke sensor 4, and a vehicle speed sensor 5 for various driving behaviors of the driver. Collect data, the data includes brake pedal force, brake pedal displacement, accelerator pedal travel and vehicle speed; input the collected sensor data into the hidden Markov model module 6 to identify the driver's intention; The human body sensor 1 of infrared technology transmits the real-time information collected and the driver's collision avoidance pedestrian intention data information predicted by the hidden Markov model module 6 to the central processing unit 7, that is, a single-chip microcomputer, and the input signal of the central processing unit 7 It is the human body sensing data signal output by the human body sensor and the driving intention data signal output by the driver intention identification double-layer hidden Markov HMM model module, and its output signal controls the buzzer to sound.

The process involved in this method includes the first offline training of the hidden Markov HMM model constructed based on offline experimental data collection, the second offline training, and the identification of driver collision avoidance based on the hidden Markov HMM model constructed based on actual driving behavior data collection. After the pedestrian intention data signal, the early warning processing is carried out through the single-chip microcomputer.

In view of the strong statistical foundation of the hidden Markov HMM model, the modular modeling method and the ability to deal with dynamic time series, the hidden Markov HMM model structure as shown in Figure 2 was constructed (the hidden Markov HMM The model includes the braking and acceleration hidden Markov HMM model and the speed classification module) to characterize the driver's intention to avoid collision with pedestrians and the driver's operation under the corresponding intention, and to identify the driving intention data signal of the behavior of the driver to avoid collision with pedestrians .

(2.1), the experimental data is collected through the vehicle information collection device, that is, the brake pedal stroke sensor 2, the brake pedal force sensor 3, the accelerator pedal stroke sensor 4 and the vehicle speed sensor 5, aiming at the possible actions taken by the driver to avoid collisions with pedestrians Driving behavior, collecting experimental data, including brake pedal force sensor data, brake pedal displacement sensor data, accelerator pedal travel sensor data and vehicle speed sensor data; The force sensor data, brake pedal displacement sensor data, and accelerator pedal travel sensor data are input into the braking and acceleration hidden Markov HMM model, and the vehicle speed sensor data is input into the speed classification module;

(2.2), in the braking and acceleration hidden Markov HMM model, construct the normal throttle release, fast release of the throttle, throttle hold, step on the accelerator, normal step on the brake, fast step on the brake, brake hold, release the brake A total of 9 multidimensional Gaussian hidden Markov HMM models about braking and acceleration, with and without pedal action. Apply the Baum-Welch algorithm to conduct offline training for 9 multi-dimensional Gaussian Hidden Markov HMMs, and iteratively optimize the parameters of each model;

(2.3) Since the collected experimental data is a long observation sequence, it needs to be processed in sections, so the vehicle speed signals in the same time period are numbered according to grades and input to the speed classification module. Described grade numbering is to classify and number the speed according to the size of the speed, for example: the size of the speed is 60km/h, and the grade of the speed is exactly 6.

(2.4), use the obtained two-dimensional driving behavior identification result string (braking and acceleration identification result string and vehicle speed identification result string) of the hidden Markov HMM model as the observation sequence of the driving intention identification hidden Markov HMM model , conduct off-line training and optimization on the hidden Markov model of driver collision avoidance pedestrian intention recognition, and obtain two hidden Markov HMM models of driver collision avoidance pedestrian intention: acceleration intention hidden Markov HMM model and braking parking intention hidden Markov model. Markov HMM model.

(2.5) Input the brake pedal force sensor signal, brake pedal displacement sensor signal, and accelerator pedal stroke sensor signal collected in real time under actual driving conditions into the braking and acceleration hidden Markov HMM model, and input the vehicle speed sensor signal In the speed classification module, refer to steps (2.1) and (2.2), conduct offline training and optimization of 9 multi-dimensional Gaussian hidden Markov HMMs, identify the driver's operation, and obtain the two-dimensional driving behavior hidden Markov HMM After the identification result string (braking, acceleration, and vehicle speed) is formed into the observation sequence string, it is sent to the hidden Markov HMM model of the driver’s collision avoidance pedestrian intention in step (2.4), and the Forward-Backward algorithm is applied to calculate two The multidimensional discrete hidden Markov HMM model produces the possibility of the observation sequence, and the model with the largest likelihood is selected as the driving intention data signal;

As shown in Figure 3, the patent of the present invention is applied to the HMM training process of the driver's braking and acceleration intention recognition method for pedestrian anti-collision warning, including the following steps:

(3.1) The experimental data is collected through the vehicle information collection device, that is, the brake pedal travel sensor, brake pedal force sensor, accelerator pedal travel sensor and vehicle speed sensor. The driving behavior that the driver may take to avoid collisions with pedestrians is used to collect experimental data. Data, including brake pedal force, brake pedal displacement, accelerator pedal travel and vehicle speed sensor data; after processing a long section of experimental data collected in sections, the brake pedal force, brake pedal displacement, accelerator pedal travel sensor data Input to the braking and acceleration hidden Markov HMM model, and input the vehicle speed to the speed classification module;

(3.2) In the braking and acceleration hidden Markov HMM model module, construct normal throttle release, fast release of throttle, throttle hold, step on the accelerator, normal step on the brake, fast step on the brake, brake hold, release There are 9 multidimensional Gaussian hidden Markov HMM models about braking and acceleration for braking and pedal no action. Apply the Baum-Welch algorithm to conduct offline training for 9 multi-dimensional Gaussian Hidden Markov HMMs, and iteratively optimize the parameters of each model;

(3.3) Since the collected experimental data is a long observation sequence, it needs to be processed in sections, so the vehicle speed signals in the same time period are numbered according to grades and input to the speed classification module. Described grade numbering is to classify and number the speed according to the size of the speed, for example: the size of the speed is 60km/h, and the grade of the speed is exactly 6.

(3.4), the collection of experimental data Through the vehicle information collection device, according to the different driving intentions of the driver to avoid collision with pedestrians, that is, acceleration and braking to stop, the experimental data is collected, including brake pedal force, brake pedal displacement, accelerator pedal travel and Vehicle speed sensor data;

(3.5) Input the brake pedal force sensor data, brake pedal displacement sensor data, and accelerator pedal stroke sensor data into the braking and acceleration hidden Markov HMM model, and input the vehicle speed sensor data into the speed classification module; apply The Forward-Backward algorithm calculates the likelihood of newly collected driving behavior sensor data relative to nine driving behavior multidimensional Gaussian hidden Markov HMM models, and selects the model with the largest likelihood as the driver's driving behavior identification result;

(3.6). The two-dimensional identification result string (braking, acceleration, and vehicle speed) of the driving behavior hidden Markov HMM obtained is used as the observation sequence of the hidden Markov HMM model for driver collision avoidance and pedestrian intention identification. The Hidden Markov Model of Pedestrian Intent Identification for Driver Collision Avoidance is trained and optimized offline, and two Hidden Markov Models of Driver Intent are obtained: Acceleration Hidden Markov HMM Model and Braking Parking Hidden Markov HMM Model;

(3.7), input the real-time collected brake pedal force sensor data, brake pedal displacement sensor data, and accelerator pedal stroke sensor signals into the braking and acceleration hidden Markov HMM model, and input the vehicle speed into the speed classification module, Referring to steps (1) and (2), conduct off-line training and optimization on 9 multidimensional Gaussian HMMs, identify the driver's operation, and obtain the two-dimensional identification result string of the driving behavior hidden Markov HMM (braking and acceleration and vehicle speed ), after forming the observation sequence string, send it into two driver intention identification hidden Markov HMM models, and apply the Forward-Backward algorithm to calculate the possibility of the observation sequence generated by the two driver intention multi-dimensional discrete hidden Markov HMM models respectively. The model with the highest likelihood is selected as the driver's intention data.

As shown in Figure 4, the data is collected through the vehicle information collection device, that is, the brake pedal travel sensor 2, the brake pedal force sensor 3, the accelerator pedal travel sensor 4 and the vehicle speed sensor 5, aiming at the actions that the driver may take to avoid collisions with pedestrians. Driving behavior, collect data, including brake pedal force, brake pedal displacement, accelerator pedal travel and vehicle speed; input the collected sensor data into the hidden Markov model module 6 to identify the driver’s intention to avoid collisions and pedestrians to accelerate or Brake to stop, that is, the driver may mistakenly step on the accelerator pedal to speed up or brake to stop to avoid pedestrians; the infrared technology-based human body sensor 1 located at the front of the car combines the real-time information collected with the hidden Markov model module 6 to predict The driver's intention information is transmitted to the central processing unit 7, that is, the single-chip microcomputer, and the central processing unit 7 recognizes the driver's collision avoidance pedestrian intention data signal and then performs data analysis in combination with the intention information and the infrared human body sensing data signal, and makes different decisions. Early warning processing. The data analysis process includes: if the central processing unit 7 analyzes the human body sensing data signal and the result is that there is no pedestrian in front, then no matter what the result of the driving intention data signal is, the central processing unit will not send a trigger signal to control the sound signal warning device 8 Carry out early warning; If the central processing unit 7 analyzes the human body induction data signal result that there are pedestrians ahead and the driving intention data signal is a braking parking intention signal; the central processing unit does not send a trigger signal to control the sound-saving signal warning device 8 to carry out early warning; if The central processing unit 7 analyzes the human body sensing data signal and the result is that there are pedestrians ahead and the driving intention data signal is an acceleration intention signal, that is, when the driver accidentally steps on the accelerator pedal to accelerate and pass, the central processing unit 7 sends a trigger signal to control the sound signal warning device 8 Carry out early warning, so as to achieve the goal that the driver can respond in time to protect pedestrians.

The above is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto, any person familiar with the technical field within the technical scope disclosed in the present invention, according to the technical solution of the present invention Any equivalent replacement or change of the inventive concepts thereof shall fall within the protection scope of the present invention.

Claims (2)

1. based on chaufeur braking and the pedestrian's anticollision method for early warning that accelerates intention identification, it is characterized in that: comprise the steps:

I, collection observed data carry out an off-line training of hidden Markov HMM model:

For chaufeur, be the driving behavior that collision prevention pedestrian may take, gather observed data, described observed data comprises brake pedal force data, brake pedal displacement data, Das Gaspedal run-length data and vehicle speed data; After the long section observed data staging treating gathering, brake pedal force data, brake pedal displacement data, Das Gaspedal run-length data are input in braking and acceleration hidden Markov HMM model, vehicle speed data is input in velocity stages module; Braking with accelerate in Hidden Markov HMM model, build normal loose throttle, fast accelerator releasing, throttle keeps, step on the throttle, normally step on braking, fast step on braking, braking maintenance, take-off the brake and pedal attonity totally 9 about braking and multidimensional Gauss's hidden Markov HMM models of acceleration; Application Baum-Welch algorithm, carries out off-line training, each model parameter of iteration optimization to described 9 multidimensional Gauss Hidden Markov HMM models; The vehicle speed signal of same time period is pressed to Grade numbers, be input to velocity stages module;

II, collection observed data carry out the secondary off-line training of Hidden Markov HMM model: for chaufeur, be the driving behavior that collision prevention pedestrian may take, gather observed data, described observed data comprises brake pedal force data, brake pedal displacement data, Das Gaspedal run-length data and vehicle speed data; Brake pedal force data, brake pedal displacement data, Das Gaspedal run-length data are input to braking again and accelerate and, in hidden Markov HMM model, vehicle speed data is input in velocity stages module again; Application Forward-Backward algorithm calculates respectively the driving behavior observed data that newly collects with respect to the likelihood score of 9 driving behavior multidimensional Gauss Hidden Markov HMM models described in step I, selects the model of likelihood score maximum as chaufeur driving behavior identification result; And the two-dimentional driving behavior identification result string-braking of this Hidden Markov HMM model and acceleration identification result string and speed of a motor vehicle identification result string, observation sequence as driver intention identification hidden Markov HMM model, chaufeur collision prevention pedestrian is intended to identification HMM and carries out off-line training and optimization, obtain 2 chaufeur collision prevention pedestrians and be intended to Hidden Markov HMM model: accelerate Hidden Markov HMM model and stopping-down Hidden Markov HMM model;

III, pick out chaufeur collision prevention pedestrian and be intended to data-signal: by the brake pedal force sensor signals of Real-time Collection, brake pedal displacement transducer signal, Das Gaspedal stroke sensor signal is input to braking and accelerates in Hidden Markov HMM model, car speed sensor signal is input in velocity stages module, 9 driving behavior multidimensional Gauss Hidden Markov HMM models are carried out to off-line training and optimization, pick out driver's operation, obtain the two-dimentional identification result string-braking and acceleration identification result string and speed of a motor vehicle identification result string of driving behavior Hidden Markov HMM model, form after observation sequence string, send into 2 Hidden Markov HMM models of driver intention identification, application Forward-Backward algorithm, calculate respectively the possibility that 2 Multidimensional Discrete Hidden Markov HMM models produce this observation sequence, select the model of likelihood score maximum as driving intention data-signal,

IV, pick out chaufeur collision prevention pedestrian and be intended in conjunction with infrared human body induction data-signal, carry out data analysis after data-signal, and make different early warning and process: the human inductor in automotive front setting based on infrared technology, for Real-time Collection human body sensing data-signal; In automobile mounted system, be provided for carrying out real-time data analysis and controlling the central processing unit that aud. snl. alarming device carries out early warning according to human body sensing data-signal and driving intention data-signal.

2. according to claim 1 based on chaufeur braking and the pedestrian's anticollision method for early warning that accelerates intention identification, it is characterized in that: described data analysis process comprises: if central processing unit analyzes human body sensing data-signal result, be that the place ahead does not exist pedestrian, no matter which kind of result driving intention data-signal is, central processing unit is not all sent energizing signal control aud. snl. alarming device and carried out early warning; If it is that the place ahead is while existing pedestrian and driving intention data-signal to be stopping-down intention signal that central processing unit analyzes human body sensing data-signal result; Central processing unit is not sent energizing signal control aud. snl. alarming device and is carried out early warning; If it is that the place ahead exists pedestrian and driving intention data-signal for accelerating intention signal that central processing unit analyzes human body sensing data-signal result; be that chaufeur is while can error stepping on accelerator pedal accelerating to pass through; central processing unit is sent energizing signal control aud. snl. alarming device and is carried out early warning, thereby reaches the object that chaufeur can make a response to protect pedestrian in time.

Images